https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126410
Bug ID: 126410
Summary: vectorizer aligns 16-byte array to 64byte
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: hjl.tools at gmail dot com
Target Milestone: ---
[hjl@gnu-tgl-3 pr126403]$ cat foo.c
#define N 4
typedef struct { float values[N]; } vector;
extern vector v;
extern void foo (vector *);
void
func (vector a, vector b)
{
vector r;
for (int i = 0; i < N; i++)
r.values[i] = a.values[i] * b.values[i];
foo (&r);
}
[hjl@gnu-tgl-3 pr126403]$ gcc -O2 -S foo.c -march=x86-64-v4
[hjl@gnu-tgl-3 pr126403]$ cat foo.s
.file "foo.c"
.text
.p2align 4
.globl func
.type func, @function
func:
.LFB0:
.cfi_startproc
vpextrq $1, %xmm4, %rdx
vmovdqa %xmm0, %xmm4
pushq %rbp
.cfi_def_cfa_offset 16
.cfi_offset 6, -16
vpinsrq $1, %rdx, %xmm4, %xmm4
vpextrq $1, %xmm5, %rdx
vmovdqa %xmm2, %xmm5
vpinsrq $1, %rdx, %xmm5, %xmm5
vpunpcklqdq %xmm1, %xmm4, %xmm4
vpunpcklqdq %xmm3, %xmm5, %xmm5
movq %rsp, %rbp
.cfi_def_cfa_register 6
andq $-64, %rsp <<< Not needd.
vmulps %xmm5, %xmm4, %xmm1
subq $64, %rsp
movq %rsp, %rdi
vmovaps %xmm1, (%rsp)
call foo
leave
.cfi_def_cfa 7, 8
ret
.cfi_endproc
.LFE0:
.size func, .-func
.ident "GCC: (GNU) 16.1.1 20260703 (Red Hat 16.1.1-4)"
.section .note.GNU-stack,"",@progbits
[hjl@gnu-tgl-3 pr126403]$
Breakpoint 1, ensure_base_align (dr_info=0x4f58288)
at /export/gnu/import/git/gitlab/x86-gcc/gcc/tree-vect-stmts.cc:7189
7189 DR_TARGET_ALIGNMENT (dr_info).to_constant () * BITS_PER_UNIT;
Missing rpms, try: dnf --enablerepo='*debug*' install
isl-debuginfo-0.16.1-24.0.fc44.x86_64 libmpc-debuginfo-1.4.1-1.fc44.x86_64
mpfr-debuginfo-4.2.2-3.fc44.x86_64 gmp-debuginfo-6.3.0-5.0.fc44.x86_64
zlib-ng-compat-debuginfo-2.3.3-3.fc44.x86_64
libzstd-debuginfo-1.5.7-5.fc44.x86_64 glibc-debuginfo-2.43-7.0.fc44.x86_64
(gdb) next
7191 if (decl_in_symtab_p (base_decl))
(gdb) p align_base_to
$1 = 512
(gdb)
Why does DR_TARGET_ALIGNMENT (dr_info).to_constant () return 64 bytes?
